• Compute a set of features, and find a simple distance between edgels based on those features.
  ** Could learn a distance if the simple solution is insufficient, but we'd need some training data.

Features ideas:

• Image:
  ** raw intensity, moments, histogram
• Shape:
  ** raw image coordinates, curvature

In particular, what 'window' is used to compute direction normal to edgel.

Should look across crack.

Potential pitfalls:

• Normal direction may not be correct for search
  *expanding crack
  *
• Do we per-determine different 'sides' of the crack? or let the search determine 'sides'?
  ** per-determination could be especially bad for "sub-cracks"